In the previous section, it was stated that neural networks employ weighted sums followed by nonlinear transformations to determine what values are used in the next layer. While this is generally true for feed-forward networks, this does not have to be true for all networks; different mathematical functions can be used apart from weighted sums to determine the next layer's input. In a different type of neural network, called a CNN, a convolution operation is used to determine the next layer's input. Typically, convolution operations are immediately followed by pooling operations (such as max pooling, in which the largest value is chosen from a discrete area to pass on to the next layer).

Networks that employ these types of operations are well-suited to data that is regularly sampled, such as images or time series data. Therefore, convolutional neural networks are popular in healthcare for processing data taken from pathology slides, radiology scans, and other images, and for detecting various diseases from them.

Pathology is a branch of medicine concerned with the evaluation of cross-sectional microscopic slides taken from human tissue samples. Examination of the slides is often done to classify tissue as cancerous or non cancerous. A pathologist sometimes has to examine very large images to look for any signs of cancerous tissue, which is very time-consuming and susceptible to errors. A recent study by Google, Inc. sought to determine whether a convolutional neural network could do a better job than pathologists at detecting breast cancer in lymph node tissue (Liu et al., 2017). For the study, slides as large as 100,000 x 100,000 pixels were used. The model was able to achieve an AUC between 0.965 and 0.986, while the pathologists achieved an AUC of 0.966. More significantly, the model produced its classification virtually instantaneously, while the human pathologist took 30 hours! This is an example of how artificial intelligence can be combined with human breadth and depth of knowledge to enhance cancer detection from pathology images. Similar types of studies can be performed on radiology scans as well.